Skip to main content

Prometheus

Admin → Clusters → Prometheus, one setting per cluster. Without it a cluster's charts use the in-memory samples (24 hours at most). Any Prometheus-compatible query API works: Prometheus, kube-prometheus-stack, Thanos, Mimir, Cortex, VictoriaMetrics, Azure Monitor managed Prometheus.

ModeWhenWhat it needs
In the cluster (API server proxy)Prometheus is only reachable inside the cluster (a ClusterIP Service, the usual kube-prometheus-stack setup)namespace, Service, port (Find in cluster fills them). OpenConsole goes through the Kubernetes API with the cluster's credentials, so no network path or Prometheus password is needed, but the ServiceAccount needs services/proxy get on that Service: uncomment the rule in deploy/rbac.yaml and name only your Service.
Direct URLPrometheus has an address OpenConsole can reach (ingress, Thanos / Mimir, managed service)URL (a path prefix is kept); authentication: none, basic, bearer token, Azure AD client credentials (scope https://prometheus.monitor.azure.com/.default; for Azure Monitor managed Prometheus), or a client certificate; extra headers (e.g. X-Scope-OrgID); CA certificate or skip verification.

Passwords, tokens, client secrets, client keys and header values are stored encrypted and never returned; the form shows a mask, and saving the mask keeps the stored value. A change is audited as cluster.update with changed: [prometheus].

If one Prometheus holds several clusters, set the label that tells them apart (cluster, cluster_name, …) and this cluster's value; Test connection lists the values it finds. Without it the charts add the clusters up.

Test connection checks, in order, and stops at the first failure: reachable (DNS / TLS / timeout told apart), allowed in, a Prometheus API, container metrics for this cluster, the cluster filter, and whether cAdvisor is scraped twice (harmless: OpenConsole takes the maximum per container).

Queries use the cAdvisor metrics container_cpu_usage_seconds_total and container_memory_working_set_bytes (no kube-state-metrics needed; requests and limits come from the pod spec), max by container against double scrapes, and a rate window of at least 5 minutes so 60 s scrape intervals work. A deployment's chart includes the pods of its earlier revisions. When Prometheus fails, the chart falls back to the in-memory samples and says so.